From Prompting to Composing: A Spatial Canvas Interface for Poster Generation

Yitong Wang; Fangyun Wei; Jinjing Zhao; Sirui Zhang; Hongyang Zhang; Dong Chen; Bo Dai; Yan Lu

cs.CVOct 8, 2026

From Prompting to Composing: A Spatial Canvas Interface for Poster Generation

Authors: Yitong Wang, Fangyun Wei, Jinjing Zhao, Sirui Zhang, Hongyang Zhang, Dong Chen, Bo Dai, Yan Lu

Organizations: Fudan University · Microsoft Research · The University of Sydney · USTC · University of Waterloo · The University of Hong Kong

Abstract

Text prompting is an indirect interface for poster generation, requiring users to encode inherently two-dimensional composition intent into a one-dimensional sequence of words. We introduce a Spatial Canvas Interface that enables users to directly compose generation intent in space through four complementary binding types: semantic, identity, text, and pixel, together with Text Specifications for individual elements and global appearance. Based on this interface, we develop Compo, a poster generation model adapted from a pretrained image editing model to understand Spatial Canvas inputs and Text Specifications. Compo supports both direct inference, where users explicitly construct the canvas, and agentic mode, where a high-level request is automatically translated into a planned Spatial Canvas. To train Compo, we develop a scalable pipeline that automatically constructs supervision data for different binding types and their combinations, enabling efficient adaptation without training a specialized poster generator from scratch. We further introduce a benchmark that evaluates adherence to individual binding types and their joint composition. Experiments show that Compo achieves stronger compositional controllability than both general-purpose image generation models and dedicated poster generation systems while maintaining high visual quality. By decoupling intent specification from visual generation, our work shifts poster generation from prompting toward composing.

Figures & tables

Figure 1: Instead of describing an entire poster through a single text prompt , our Spatial Canvas Interface allows users to directly compose generation intent in two-dimensional space. The interface supports four complementary forms of binding: Semantic Binding specifies what visual content should be generated in a designated region; Identity Binding associates a reference image with a target region to preserve its visual identity; Text Binding places specified text at a desired location; and Pixel Binding directly anchors visual content that should be preserved. The accompanying Text Specifications further provide detailed instructions for individual bindings, such as the desired appearance, action, or content of each bound element, together with global specifications for the background and style. By jointly expressing spatial structure on the canvas and detailed generation requirements in text, our interface provides an explicit and structured specification of the target poster, enabling generation through composition rather than text description alone.

Figure 2: Training data generation pipeline. We first create diverse posters through structured planning and information-preserving rewriting with GPT-5.5 ( OpenAI, 2026c ) , followed by rendering with Qwen-Image-2512 ( Wu et al., 2025a ) , Z-Image ( Cai et al., 2025 ) , and ERNIE-Image ( Liu et al., 2026 ) . We then automatically construct paired Spatial Canvases and Text Specifications through instance segmentation, visual asset processing, and text recognition, providing scalable supervision for training Compo.

Figure 3: Architecture and training pipeline of Compo. Compo is trained in two stages. During supervised training, the Spatial Canvas and Text Specifications are jointly encoded with the target poster to adapt a pretrained MMDiT ( Peebles and Xie, 2023 ; Esser et al., 2024 ) through LoRA ( Hu et al., 2021 ) . We then apply DiffusionNFT ( Zheng et al., 2026 ) for reinforcement learning with complementary rewards for text accuracy, instance similarity, and layout adherence.

Figure 4: Poster generation using Compo with Agentic Mode. Given a high-level user request, a GPT-5.6-based agent extracts visual elements and text, plans their spatial layout, assigns appropriate binding types, and retrieves reference images when needed. It then constructs a planned Spatial Canvas and Text Specifications, which are passed to Compo to generate the final poster.

Models	Pixel Binding					Identity Binding					Semantic Binding		Text Binding			Overall
Models	DINO-I	CLIP-I	M-DINO	SSIM	IoU	DINO-I	CLIP-I	M-DINO	CLIP-T	IoU	CLIP-T	IoU	NED	Sen. Acc.	IoU	IoU	Avg. Score
General Text-and-Image-to-Image Models
Step1X-Edit ( Liu et al., 2025 )	0.426	0.481	0.327	0.493	0.473	0.600	0.645	0.488	0.167	0.563	0.012	0.056	0.060	0.020	0.270	0.317	0.339
GLM-Image ( Z.ai, 2026 )	0.535	0.580	0.441	0.591	0.571	0.663	0.699	0.554	0.185	0.606	0.019	0.110	0.030	0.012	0.230	0.328	0.389
BAGEL ( Deng et al., 2025 )	0.274	0.311	0.207	0.321	0.321	0.534	0.580	0.428	0.156	0.511	0.006	0.045	0.007	0.000	0.184	0.228	0.259
ChronoEdit-14B ( Wu et al., 2025b )	0.575	0.629	0.479	0.647	0.638	0.697	0.733	0.596	0.192	0.631	0.011	0.055	0.004	0.000	0.260	0.350	0.410

Table 1: We compare general text-and-image-to-image models, dedicated poster generation models, and our Compo variants across pixel, identity, semantic, and text binding metrics. Higher values indicate better performance. The best and second-best results are highlighted in orange and blue, respectively. “NED”: normalized edit distance. “Sen. Acc.”: sentence accuracy.

Figure 5: Visual comparison with representative baselines. We compare Compo with dedicated poster generation methods (PosterMaker ( Gao et al., 2025 ) , PosterOmni ( Chen et al., 2026d ) , and CreatiDesign ( Zhang et al., 2026 ) ) and the best-performing general text-and-image-to-image baseline in Table 1 , HiDream-O1-Image ( Cai et al., 2026 ) . Transparent pink boxes highlight noticeable errors. Additional visualizations are provided in Appendix C.2 .

Reward Setting	Pixel Binding					Identity Binding					Semantic Binding		Text Binding			Overall
Reward Setting	DINO-I	CLIP-I	M-DINO	SSIM	IoU	DINO-I	CLIP-I	M-DINO	CLIP-T	IoU	CLIP-T	IoU	NED	Sen. Acc.	IoU	IoU	Avg. Score
Full Reward (Default)	0.735	0.776	0.642	0.793	0.781	0.795	0.891	0.757	0.261	0.904	0.259	0.828	0.970	0.940	0.944	0.885	0.752
w/o Text Reward	0.752	0.789	0.666	0.801	0.790	0.811	0.901	0.776	0.262	0.928	0.275	0.884	0.878	0.778	0.895	0.876	0.746
w/o Ins. Sim. Reward	0.725	0.769	0.639	0.788	0.776	0.782	0.879	0.740	0.255	0.877	0.239	0.753	0.958	0.924	0.942	0.866	0.736
w/o Layout Reward	0.732	0.771	0.644	0.787	0.779	0.799	0.892	0.760	0.260	0.902	0.263	0.830	0.968	0.930	0.933	0.878	0.750
w/o All Rewards	0.716	0.763	0.619	0.785	0.775	0.780	0.878	0.733	0.257	0.893	0.226	0.720	0.924	0.886	0.916	0.849	0.725

Table 2: Effects of different RL reward components. We ablate the text, instance similarity (“Ins. Sim.”), and layout rewards individually and jointly. The best and second-best results are highlighted in orange and blue, respectively. “NED”: normalized edit distance. “Sen. Acc.”: sentence accuracy.

Sen. Acc. Weight	Text Binding
Sen. Acc. Weight	NED	Sen. Acc.	IoU
0.75 (Default)	0.970	0.940	0.944
0.0	0.971	0.916	0.945
0.25	0.966	0.920	0.944
0.5	0.966	0.928	0.945
1.0	0.969	0.938	0.938

Table 3: Effects of different sentence accuracy reward weights. Smaller weights tolerate minor text rendering errors, while larger weights over-rely on the sparse sentence accuracy during the early stage of RL post-training. The best and second-best results are highlighted in orange and blue, respectively. “NED”: normalized edit distance. “Sen. Acc.”: sentence accuracy.

LoRA Target Modules	Pixel Binding					Identity Binding					Semantic Binding		Text Binding			Overall
LoRA Target Modules	DINO-I	CLIP-I	M-DINO	SSIM	IoU	DINO-I	CLIP-I	M-DINO	CLIP-T	IoU	CLIP-T	IoU	NED	Sen. Acc.	IoU	IoU	Avg. Score
Both Branches (Default)	0.716	0.763	0.619	0.785	0.775	0.780	0.878	0.733	0.257	0.893	0.226	0.720	0.924	0.886	0.916	0.849	0.725
Image Branch Only	0.712	0.764	0.623	0.787	0.777	0.772	0.873	0.731	0.256	0.881	0.226	0.710	0.913	0.878	0.906	0.841	0.721
Text Branch Only	0.714	0.764	0.626	0.786	0.775	0.770	0.868	0.722	0.253	0.882	0.224	0.711	0.926	0.891	0.916	0.846	0.722

Table 4: Effects of different LoRA target modules. The best and second-best results are highlighted in orange and blue, respectively. “NED”: normalized edit distance. “Sen. Acc.”: sentence accuracy.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Models	Pixel Binding					Identity Binding					Semantic Binding		Text Binding			Overall
Models	DINO-I	CLIP-I	M-DINO	SSIM	IoU	DINO-I	CLIP-I	M-DINO	CLIP-T	IoU	CLIP-T	IoU	NED	Sen. Acc.	IoU	IoU	Avg. Score
Compo-LongCat-Base	0.716	0.763	0.619	0.785	0.775	0.780	0.878	0.733	0.257	0.893	0.226	0.720	0.924	0.886	0.916	0.849	0.725
Compo-Qwen-Base	0.720	0.765	0.631	0.785	0.772	0.791	0.891	0.758	0.259	0.918	0.217	0.716	0.907	0.858	0.906	0.845	0.726
Compo-LongCat	0.735	0.776	0.642	0.793	0.781	0.795	0.891	0.757	0.261	0.904	0.259	0.828	0.970	0.940	0.944	0.885	0.752
Compo-Qwen	0.734	0.770	0.649	0.782	0.773	0.810	0.910	0.781	0.263	0.932	0.262	0.850	0.940	0.868	0.934	0.886	0.751

Appendix

Table 5: We compare our Compo models before and after RL post-training across pixel, identity, semantic, and text binding metrics. Higher values indicate better performance. The best and second-best results are highlighted in orange and blue, respectively. “NED”: normalized edit distance. “Sen. Acc.”: sentence accuracy.

Figure 6: Additional visual comparison with representative baselines (Part 1). We compare Compo with dedicated poster generation methods (PosterMaker ( Gao et al., 2025 ) , PosterOmni ( Chen et al., 2026d ) , and CreatiDesign ( Zhang et al., 2026 ) ) and the best-performing general text-and-image-to-image baseline in Table 1 , HiDream-O1-Image ( Cai et al., 2026 ) . Transparent pink boxes highlight noticeable errors.

Figure 7: Additional visual comparison with representative baselines (Part 2). We compare Compo with dedicated poster generation methods (PosterMaker ( Gao et al., 2025 ) , PosterOmni ( Chen et al., 2026d ) , and CreatiDesign ( Zhang et al., 2026 ) ) and the best-performing general text-and-image-to-image baseline in Table 1 , HiDream-O1-Image ( Cai et al., 2026 ) . Transparent pink boxes highlight noticeable errors.

Figure 8: Additional visualizations at various resolutions across diverse trending themes and visual elements. Our Compo models can generate visually coherent and plausible poster designs based on diverse trending themes and visual elements that are unseen during training, demonstrating robust generalization and practical applicability.

Explore similar work

Any2Poster: Any-Source Poster Generation Across Modalities and Domains
Jun 1, 2026Amogh Vinaykumar, Aiden Li, Suozhi Huang +1Image Generation Multimodal Generation
Flower Mound High School · University College London · Princeton University
simpleposter: a simple baseline for product poster generation
May 9, 2026Benlei Cui, Fangao Zeng, Weitao Jiang +6Image Generation Image Inpainting
1Alibaba Group · 2Taobao & Tmall Group of Alibaba
PosterHarness: Turning Scientific Poster Generation into an Auditable Instruction-Following Benchmark
Jul 3, 2026Tianyi Yang, Dawei Fu, Youpeng Wu +5AI-Assisted Scientific Research Image Generation
School of Physics, Peking University

Jun 1, 2026, cs.CVJ/K move · Enter open · S save

Any2Poster: Any-Source Poster Generation Across Modalities and Domains

Amogh Vinaykumar, Aiden Li, Suozhi Huang, Shilong Liu

Flower Mound High School · University College London · Princeton University

Image Generation Multimodal Generation Automated Evaluation

Open paper AskPDF arXiv

Visual posters are a compact medium for communicating dense information, yet progress on automatic poster generation remains difficult to measure because existing evaluations are often restricted to paper-only inputs, narrow domains, or surface-level visual similarity. We introduce Any2Poster Bench, a benchmark for any-source poster generation that evaluates systems across eight input modalities--PDFs, URLs, PPTX, DOCX, Markdown, LaTeX, notebooks, and videos--and five content domains. Any2Poster Bench pairs each source with quiz-based probes of verbatim factual retention and interpretive understanding, together with VLM-based judgments of visual quality, layout, readability, content completeness, and logical flow, enabling reproducible assessment of both information fidelity and visual communication. To instantiate and validate this benchmark, we further present Any2Poster Agent, an end-to-end reference agent that parses heterogeneous sources, organizes salient content, plans poster layouts, renders posters, and iteratively refines them using visual feedback. On Any2Poster Bench, Any2Poster Agent achieves 87.25% average accuracy across input modalities and 87.28% across content domains. On PaperQuiz-style evaluation, where prior paper-to-poster agents are directly comparable, Any2Poster Agent improves over PosterAgent-4o from 51.06-51.33% to 72.58% overall accuracy and from 116-121 to 145.16 in density-augmented score. Together, Any2Poster Bench and Any2Poster Agent provide a reusable evaluation resource and a competitive baseline for studying multimodal, domain-general poster generation.

Abstract

Explore similar work

Any2Poster: Any-Source Poster Generation Across Modalities and Domains

simpleposter: a simple baseline for product poster generation

PosterHarness: Turning Scientific Poster Generation into an Auditable Instruction-Following Benchmark